Get In Touch
AI Gaps Calculator · Brillian

What Are Your AI Gaps
Costing You?

For software teams shipping AI features. See which production-readiness gaps you have and what they cost you each year.

01 — Your team

clients
Contract Value — calculated
Annual revenue ÷ Customers€63K
engineers

02 — Select your gaps

How this works

What the calculator measures. For each gap you select, the model estimates three types of cost: engineering capacity lost to manual validation, rework, and maintenance; revenue at risk from increased churn and slower delivery; and roadmap opportunity cost from compounding technical debt. These are added together with diminishing returns applied — gaps share root causes, so the total is weighted rather than a straight sum.

HIGH, MED, LOW ratings show the relative severity of each gap based on how much it typically affects engineering throughput, customer trust, and commercial outcomes across B2B software teams.

Conservative, Base, and Aggressive scenarios shift both the gap weightings and the overall cap. The cap prevents implausible outputs — set at 15–28% of annual revenue and 28–45% of engineering cost depending on scenario.

Treat all figures as directional diagnostic signals, not financial forecasts.

How we calculated this

Formula. For each selected gap, the model estimates two cost buckets. Engineering cost = your team size × average cost × gap weight. Revenue cost = a combination of churn exposure (customers × contract value × churn rate × gap severity) and roadmap opportunity cost (annual revenue × delivery delay factor). The two buckets are summed with diminishing returns applied — gaps are ordered by severity; the top four (all HIGH-rated) carry full weight, the next four carry 50%, and the bottom two carry 20%.

Where the weights come from. Gap weights are Brillian's estimates, calibrated from published research — not figures cited directly in any single study. HIGH gaps carry the most weight because they affect both engineering capacity and revenue risk daily. MED gaps carry moderate weight. LOW gaps carry the least.

Revenue at risk. Revenue at risk estimates are informed by algorithm aversion research (Dietvorst 2015), AI product retention benchmarks, and software reliability impact data.

What this means for the numbers. The output is a directional estimate — the right order of magnitude, not a precise forecast. Two companies with identical inputs can have very different actual costs depending on how their AI features are built and used. The calculation is a starting point for a conversation, not a conclusion.

Research & sources

The studies below informed the gap weightings. Confidence level is noted per source.

Directly measured
Industry survey
Applied by inference
  1. Source 01
    Peng, Kalliamvakou et al. (2023)
    GitHub Copilot. N=95 controlled experiment — 55.8% faster on coding tasks. Used to calibrate the engineering productivity impact of missing evaluation and shared platform infrastructure.
    https://arxiv.org/abs/2302.06590
    Directly measured
  2. Source 02
    Noy & Zhang (2023)
    MIT randomised trial — AI reduced task time by 40%. Used alongside Peng et al. to set the engineering productivity baseline.
    https://economics.mit.edu/sites/default/files/2023-03/noy_zhang_1.pdf
    Directly measured
  3. Source 03
    Stripe (2018)
    Developer Coefficient. N=1,000+. Developers spend 42% of time on debt and maintenance. The core anchor for engineering capacity lost estimates across all gaps. Corroborated by Unqork/Morning Consult 2024 (tech debt stifles innovation at 80% of enterprises) and Stack Overflow 2024 (62% of developers cite tech debt as their greatest frustration).
    https://stripe.com/reports/developer-coefficient-2018
    Directly measured
  4. Source 04
    RAND Corporation (2025)
    AI failure study. N=65 practitioners. 80%+ of AI projects fail — twice the rate of conventional software. Used to calibrate the revenue and roadmap risk weightings for HIGH-rated gaps.
    https://www.rand.org/pubs/research_reports/RRA2680-1.html
    Directly measured
  5. Source 05
    Deloitte AI Institute (2024)
    State of GenAI in the Enterprise. N=2,773, 14 countries. 74% of advanced AI initiatives meet ROI — but two-thirds of experiments don't scale. Used to calibrate the commercial impact of gaps on expansion and retention.
    https://www.deloitte.com/us/en/about/press-room/state-of-generative-ai.html
    Industry survey
  6. Source 06
    Microsoft & LinkedIn (2024)
    Work Trend Index. N=31,000, 31 countries. 75% of knowledge workers use AI. Used to set adoption baseline and inform the trust and inconsistency gap impact estimates.
    https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part
    Industry survey
  7. Source 07
    Freshworks CX Benchmark (2024–25)
    Real deployment data. AI agents deflect 45% of support queries; response time improved 43%. Used to calibrate the revenue upside of well-scoped AI and the cost of inconsistent outputs.
    https://www.freshworks.com/How-AI-is-unlocking-ROI-in-customer-service/
    Industry survey
  8. Source 08
    Sculley et al. (2015)
    Hidden Technical Debt in ML Systems. NeurIPS 2015. Applied by inference to estimate the compounding maintenance cost of missing shared platform infrastructure and regression testing.
    https://papers.nips.cc/paper/5656
    Based on research
  9. Source 09
    DORA — State of AI-Assisted Software Development (2025)
    N=5,000 professionals. AI amplifies what is already there — strong teams get better, struggling teams get worse. Teams without robust automated testing, version control, and feedback loops saw stability decrease 7.2% after AI adoption (DORA 2024, N=39,000). Directly validates the evaluation, observability, and shared platform gap impact estimates.
    https://dora.dev/dora-report-2025/
    Industry survey
  10. Source 10
    McKinsey Global Institute (2023)
    The Economic Potential of Generative AI. One full day per week lost to information search. Applied by inference to calibrate the revenue opportunity cost of delayed or poorly governed AI features.
    https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-economic-potential-of-generative-ai
    Based on research
  11. Source 11
    Dietvorst, Simmons & Massey (2015)
    Algorithm Aversion: People Erroneously Avoid Algorithms After Seeing Them Err. Wharton School, five controlled studies. Users abandon AI systems after errors more rapidly than human experts — even when AI outperforms the human. Applied to calibrate the revenue risk multipliers for hallucination and inconsistency gaps.
    https://doi.org/10.1037/xge0000033
    Based on research
  12. Source 12
    Stack Overflow Developer Survey (2025)
    N=49,000+, 177 countries. Only 30% of developers trust AI-generated outputs — down from 40% in 2024. 66% cite "almost right but not quite" as biggest AI frustration. Positive sentiment dropped from 70%+ to 60%. Directly validates the trust and inconsistency gap impact estimates.
    https://survey.stackoverflow.co/2025/
    Industry survey
  13. Source 13
    Deloitte AI Institute — Trust in AI (2024)
    Analysis of 1,982 respondents. Trust builders — those reducing hallucinations, improving observability, ensuring explainability — were 18 percentage points more likely to achieve expected AI benefits. First direct quantified link between production infrastructure rigour and commercial outcomes.
    https://www.deloitte.com/us/en/insights/topics/risk-management/trust-in-ai.html
    Industry survey
  14. Source 14
    Shankar et al. (2024), ACM CSCW
    Field study of ML production debugging and incident handoff. Illuminates the day-to-day cost of missing observability and audit infrastructure in real teams.
    https://dl.acm.org/doi/10.1145/3653697
    Industry survey
  15. Source 15
    Ghosh et al. (2022), Microsoft / SoCC
    Cloud infrastructure study. Many production incidents do not trigger existing monitors — supports the case for purpose-built AI observability over general-purpose alerting.
    https://dl.acm.org/doi/10.1145/3542929.3563482
    Directly measured
  16. Source 16
    Eken et al. (2025), ACM Computing Surveys
    Systematic review of MLOps pain points across the literature. Provides broad evidence base for the engineering overhead estimates across multiple gap types.
    https://dl.acm.org/doi/10.1145/3747346
    Based on research
  17. Source 17
    Reichheld & Sasser (1990), Harvard Business Review
    Zero Defections: Quality Comes to Services. Original source for the finding that a 5% increase in retention can increase profits by 25–95%. Foundational reference for the revenue at risk model.
    https://hbr.org/1990/09/zero-defections-quality-comes-to-services
    Directly measured
  18. Source 18
    MIT NANDA — The GenAI Divide (2025)
    State of AI in Business 2025. Analysis of 300+ public AI deployments, 150 executive interviews, N=350 survey. 95% of enterprise GenAI pilots fail to deliver measurable financial impact despite global investment exceeding $30 billion. The core issue is not model quality but missing operational foundations. Supplements RAND 2025 in calibrating revenue risk for teams without AI infrastructure.
    https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
    Directly measured
  19. Source 19
    Grafana Labs — Observability Survey (2025)
    N=1,000+ engineering professionals. Only 3% of organisations have AI observability deployed in production. Most teams rely on general-purpose monitoring that was not designed for AI behaviour. Directly validates the observability gap and the case for purpose-built AI monitoring.
    https://grafana.com/observability-survey/2025/
    Industry survey